Computer‑use agents (CUAs) act on behalf of users in online platforms. When the platform’s incentives diverge from the user’s goal, agents can be steered away from the intended outcome. To study this, we introduce the CAVEAT benchmark, which spans nine marketplace settings and enumerates eight typical steering mechanisms. We evaluate five families of models: in a matched‑control setting without steering, agents purchase the user‑optimal product 78.6% of the time; with steering enabled the rate drops to 17.3%. Larger models and stronger reasoning improve robustness, yet failures remain common. Trajectory analysis and targeted ablations reveal three entry points for steering: distortion of user priorities, premature narrowing of the alternative set, and committing before decision‑relevant evidence is resolved. Guided by this diagnosis we build CAVEAT‑Harness, which directly addresses these failure modes and raises user‑optimal purchasing to 55.0%. Additional post‑training further benefits a smaller open‑source model. These findings establish incentive robustness as a distinct challenge for delegated agents and show that focused interventions can substantially mitigate it.
Review